Automated Vertebral Landmarks and Spinal Curvature Estimation using Non-directional Part Affinity Fields

Abstract Vertebral landmarks of posterior-anterior X-ray images can be used to determine the curvature of the spine, which is essential for the assessment of Adolescent Idiopathic Scoliosis. Recently, automatic methods are presented to assist the annotation process of vertebral landmarks. However, previous methods expect a fixed number of landmark outputs regardless of the occlusion and diversity in the X-ray images, which may result in ina0ccurate predictions in practical cases. The use of dense layers and fractional numbers with respect to image dimensions as landmark outputs also limited the accuracy of the prediction. In this paper, we propose a novel method based on fully convolutional architecture and non-directional part affinity fields to label arbitrary number of vertebral landmarks and keep track of the connection relations between them to compute the Cobb angles. Comparing with conventional methods which predict each landmark by dense layers directly, our method use image segmentation to predict the landmarks and their connection relationships with seven heatmap channels: four for corner landmarks and two for vertical center landmarks of each vertebra, and one for non-directional part affinity fields to express the pairing relations between left and right center landmarks. This design allows arbitrary number of landmarks as output, and reduces the impact of occluded landmark and malformed vertebrae on Cobb angles prediction. To evaluate our methods, mean squared error for landmarks and absolute percentage error for Cobb angles are tested. The results show that our method outperforms other methods on both metrics.